An AI agent's claim is not evidence. Red attacks, Blue defends, the Manager keeps score -
and only sandboxed test results decide what happened.
What a Lite battle can prove (Python code):
Command injection (CWE-78): input reaches a shell command.
Path traversal (CWE-22): input chooses which file gets read (../).
FileDesk has one of each planted, so you see the full attack → patch → retest loop.
Your own repo works too, but if it has neither pattern the honest result is no confirmed finding.
Repo Check scans any language, statically.
Battles actively test code: only use repositories you own or are authorised to assess.